Biomod: right whale feeding threshold models
Camille Ross
11/19/2020
Summary
Model v0.2.3 was created using wind, lag_sst, int_chl, sss for cfin. The models were produced using the ECOMON dataset(s). The models were used to project the probability of the study area of having a cfin abundance of over 10^{4} per \(m^2\), which is the right whale feeding threshold selected for this model and is assumed to indicate the formation of a patch sufficient for right whale feeding. The models were then built using the species distribution modeling package, Biomod2, which builds presence-absence models using any of 10 different algorithms. The algorithms selected were generalized additive models (GAMs), good explanatory models; boosted regression trees (BRTs), good predictive models; and random forests (RFs), highly accurate predictive models. One model was built for each month, and then projected back onto the environmental data from that month for every year between 2000 and 2017.
The models were averaged into climatologies with one climatology per month. Evaluations were compiled for each individual year and plotted by month. Finally, the study area was divided up into three regions, the Mid-Atlantic Bight (MAB), George’s Bank (GBK), and the Gulf of Maine (GOM). Actual versus predicted abundance values were plotted for each region.
Climatologies
Ensemble Climatology
The ensemble models were created using the biomod2 package in R. The ensembles consist of BRTs, GAMs, and RFs. The ensembles were used to model the right whale feeding threshold, with any abundance greater than 10^{4} cfin per \(m^2\) counted as a presence and anything below that threshold counted as an absence. A presence indicates the formation of a patch that is sufficient for right whale feeding according to the selected threshold.
Figure 1. Monthly climatological ensemble projections of GAMs, BRTs, and random forests (RFs). The climatology was created by averaging together the projections from 2000 to 2017.
GAM Climatology
The GAM models created with biomod2 were used to model the right whale feeding threshold, with any abundance greater than 10^{4} cfin per \(m^2\) counted as a presence and anything below that threshold counted as an absence. A presence indicates the formation of a patch that is sufficient for right whale feeding according to the selected threshold.
Figure 2. Monthly climatological GAM projections produced using Biomod2. The climatology was created by averaging together the projections from 2000 to 2017.
BRT Climatology
The BRT models created with biomod2 were used to model the right whale feeding threshold, with any abundance greater than 10^{4} cfin per \(m^2\) counted as a presence and anything below that threshold counted as an absence. A presence indicates the formation of a patch that is sufficient for right whale feeding according to the selected threshold.
Figure 3. Monthly climatological BRT projections produced using Biomod2. The climatology was created by averaging together the projections from 2000 to 2017.
RF Climatology
The RF models created with biomod2 were used to model the right whale feeding threshold, with any abundance greater than 10^{4} cfin per \(m^2\) counted as a presence and anything below that threshold counted as an absence. A presence indicates the formation of a patch that is sufficient for right whale feeding according to the selected threshold.
Figure 4. Monthly climatological RF projections produced using Biomod2. The climatology was created by averaging together the projections from 2000 to 2017.
Monthly ensemble projections
Monthly ensemble Biomod2 projections are displayed below for the months of May, June, July, August, and September.
April
Figure 5. Ensemble projections for the month of April from 2000 to 2017.
May
Figure 6. Ensemble projections for the month of May from 2000 to 2017.
June
Figure 7. Ensemble projections for the month of June from 2000 to 2017`.
August
Figure 8. Ensemble projections for the month of August from 2000 to 2017.
September
Figure 9. Ensemble projections for the month of September from 2000 to 2017.
Evaluations
Evaluation metrics were selected based on availability within the Biomod2 package. The area under the receiver operator characteristic curve (AUC) and the true skill statistic (TSS) were computed during the creation of the model object.
Ensemble evaluations
Figure 10. Biomod ensemble evaluations on a monthly time scale using a.) AUC and b.) TSS
GAM evaluations
Figure 11. Biomod GAM evaluations on a monthly time scale using a.) AUC and b.) TSS
BRT evaluations
Figure 12. Biomod BRT evaluations on a monthly time scale using a.) AUC and b.) TSS
RF evaluations
Figure 13. Biomod RF evaluations on a monthly time scale using a.) AUC and b.) TSS
Variable contribution
Variable contribution was saved during each model run and then reloaded and plotted on a monthly basis and standardized so the total contribution is equal to 100%. This was only done for the individual models.
GAM variable contribution
Figure 14. Biomod GAM variable contributions on a monthly time scale.
BRT variable contribution
Figure 15. Biomod BRT variable contributions on a monthly time scale.
RF variable contribution
Figure 16. Biomod RF variable contributions on a monthly time scale.
Climatological and inter-annual actual abundance vs. predicted probability of patch formation
For each model, the logged actual abundance of cfin was plotted against the predicted probability of patch formation This was done for both the training data (ECOMON) and independent data (CPR). Error bars indicate the variance. The plots are color coded by region, either Mid-Atlantic Bight (MAB), George’s Bank (GBK), or the Gulf of Maine (GOM).
Ensemble model
Training data (ECOMON)
Figure 17. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
Independent data (CPR)
Figure 18. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
GAM model
Training data (ECOMON)
Figure 19. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
Independent data (CPR)
Figure 20. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
BRT model
Training data (ECOMON)
Figure 21. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
Independent data (CPR)
Figure 22. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
RF model
Training data (ECOMON)
Figure 23. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
Independent data (CPR)
Figure 24. Actual logged abundance versus predicted probability of patch formation for cfin for a.) all 12 months and b.) all years.
Climatological and inter-annual actual probability of patch formation vs. predicted
For each model, the actual probability of a patch of cfin being sufficient for right whale feeding was plotted against the predicted probability of patch formation This was done for both the training data (ECOMON) and independent data (CPR). Error bars indicate the variance. The plots are color coded by region, either Mid-Atlantic Bight (MAB), George’s Bank (GBK), or the Gulf of Maine (GOM).
Ensemble model
Training data (ECOMON)
Figure 25. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
Independent data (CPR)
Figure 26. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
GAM model
Training data (ECOMON)
Figure 27. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
Independent data (CPR)
Figure 28. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
BRT model
Training data (ECOMON)
Figure 29. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
Independent data (CPR)
Figure 30. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
RF model
Training data (ECOMON)
Figure 31. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
Independent data (CPR)
Figure 32. Measured probability of patch versus predicted probability of patch for cfin for a.) monthly binning and b.) inter-annual binning.
Region plots for actual abundance
For each region, a plot was created comparing the actual abundance of cfin for both the training dataset (ECOMON) and an independent dataset (CPR) to the predicted probability of habitat suitability. The shaded confidence interval represents variance.
Ensemble region plots
Training data (ECOMON)
Climatological
Figure 33. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 34. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 35. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 36. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
GAM region plots
Training data (ECOMON)
Climatological
Figure 37. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 38. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 39. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 40. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
BRT region plots
Training data (ECOMON)
Climatological
Figure 41. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 42. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 43. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 44. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
RF region plots
Training data (ECOMON)
Climatological
Figure 45. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 46. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 47. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 48. Plots of actual abundance vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Region plots for binary patch threshold
For each region, a plot was created comparing whether or not the abundance data exceeded the right whale feeding threshold (1 and 0, respectively), indicating the formation of a sufficient patch, for both the training dataset (ECOMON) and an independent dataset (CPR) to the predicted probability of patch formation. The shaded confidence interval represents variance.
Ensemble region plots
Training data (ECOMON)
Climatological
Figure 49. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 50. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 51. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 52. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
GAM region plots
Training data (ECOMON)
Climatological
Figure 53. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 54. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 55. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 56. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
BRT region plots
Training data (ECOMON)
Climatological
Figure 57. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 58. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 59. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 60. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
RF region plots
Training data (ECOMON)
Climatological
Figure 61. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 62. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Independent data (CPR)
Climatological
Figure 63. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.
Inter-annual
Figure 64. Plots of actual probability of a patch exceeding the feeding threshold vs. predicted probability of a patch exceeding the feeding threshold in different regions.